Identifying layers within text-to-image models which control visual attributes can facilitate efficient model editing through closed-form updates. Recent work, leveraging causal tracing show that early Stable-Diffusion variants confine knowledge primarily to the first layer of the CLIP text-encoder, while it diffuses throughout the UNet.Extending this framework, we observe that for recent models (e.g., SD-XL, DeepFloyd), causal tracing fails in pinpointing localized knowledge, highlighting challenges in model editing. To address this issue, we introduce the concept of Mechanistic Localization in text-to-image models, where knowledge about various visual attributes (e.g., "style", "objects", "facts") can be mechanistically localized to a small fraction of layers in the UNet, thus facilitating efficient model editing. We localize knowledge using our method LocoGen which measures the direct effect of intermediate layers to output generation by performing interventions in the cross-attention layers of the UNet. We then employ LocoEdit, a fast closed-form editing method across popular open-source text-to-image models (including the latest SD-XL)and explore the possibilities of neuron-level model editing. Using Mechanistic Localization, our work offers a better view of successes and failures in localization-based text-to-image model editing. Code will be available at https://github.com/samyadeepbasu/LocoGen.
This study presents a targeted model editing analysis focused on the latest large language model, Llama-3. We explore the efficacy of popular model editing techniques - ROME, MEMIT, and EMMET, which are designed for precise layer interventions. We identify the most effective layers for targeted edits through an evaluation that encompasses up to 4096 edits across three distinct strategies: sequential editing, batch editing, and a hybrid approach we call as sequential-batch editing. Our findings indicate that increasing edit batch-sizes may degrade model performance more significantly than using smaller edit batches sequentially for equal number of edits. With this, we argue that sequential model editing is an important component for scaling model editing methods and future research should focus on methods that combine both batched and sequential editing. This observation suggests a potential limitation in current model editing methods which push towards bigger edit batch sizes, and we hope it paves way for future investigations into optimizing batch sizes and model editing performance.
Since the rapid development of Large Language Models (LLMs) has achieved remarkable success, understanding and rectifying their internal complex mechanisms has become an urgent issue. Recent research has attempted to interpret their behaviors through the lens of inner representation. However, developing practical and efficient methods for applying these representations for general and flexible model editing remains challenging. In this work, we explore how to leverage insights from representation engineering to guide the editing of LLMs by deploying a representation sensor as an editing oracle. We first identify the importance of a robust and reliable sensor during editing, then propose an Adversarial Representation Engineering (ARE) framework to provide a unified and interpretable approach for conceptual model editing without compromising baseline performance. Experiments on multiple tasks demonstrate the effectiveness of ARE in various model editing scenarios. Our code and data are available at https://github.com/Zhang-Yihao/Adversarial-Representation-Engineering.
While current large language models (LLMs) perform well on many knowledge-related tasks, they are limited by relying on their parameters as an implicit storage mechanism. As a result, they struggle with memorizing rare events and with updating their memory as facts change over time. In addition, the uninterpretable nature of parametric memory makes it challenging to prevent hallucination. Model editing and augmenting LLMs with parameters specialized for memory are only partial solutions. In this paper, we introduce MemLLM, a novel method of enhancing LLMs by integrating a structured and explicit read-and-write memory module. MemLLM tackles the aforementioned challenges by enabling dynamic interaction with the memory and improving the LLM's capabilities in using stored knowledge. Our experiments indicate that MemLLM enhances the LLM's performance and interpretability, in language modeling in general and knowledge-intensive tasks in particular. We see MemLLM as an important step towards making LLMs more grounded and factual through memory augmentation. The project repository is publicly available at https://github.com/amodaresi/MemLLM
How does the internal computation of a machine learning model transform inputs into predictions? In this paper, we introduce a task called component modeling that aims to address this question. The goal of component modeling is to decompose an ML model's prediction in terms of its components -- simple functions (e.g., convolution filters, attention heads) that are the "building blocks" of model computation. We focus on a special case of this task, component attribution, where the goal is to estimate the counterfactual impact of individual components on a given prediction. We then present COAR, a scalable algorithm for estimating component attributions; we demonstrate its effectiveness across models, datasets, and modalities. Finally, we show that component attributions estimated with COAR directly enable model editing across five tasks, namely: fixing model errors, ``forgetting'' specific classes, boosting subpopulation robustness, localizing backdoor attacks, and improving robustness to typographic attacks. We provide code for COAR at https://github.com/MadryLab/modelcomponents .
We investigate the mechanisms of factual recall in the Mamba state space model. Our work is inspired by previous findings in autoregressive transformer language models suggesting that their knowledge recall is localized to particular modules at specific token locations; we therefore ask whether factual recall in Mamba can be similarly localized. To investigate this, we conduct four lines of experiments on Mamba. First, we apply causal tracing or interchange interventions to localize key components inside Mamba that are responsible for recalling facts, revealing that specific components within middle layers show strong causal effects at the last token of the subject, while the causal effect of intervening on later layers is most pronounced at the last token of the prompt, matching previous findings on autoregressive transformers. Second, we show that rank-one model editing methods can successfully insert facts at specific locations, again resembling findings on transformer LMs. Third, we examine the linearity of Mamba's representations of factual relations. Finally we adapt attention-knockout techniques to Mamba in order to dissect information flow during factual recall. We compare Mamba directly to a similar-sized autoregressive transformer LM and conclude that despite significant differences in architectural approach, when it comes to factual recall, the two architectures share many similarities.
Addressing the issues of hallucinations and outdated knowledge in large language models is critical for their reliable application. Model Editing presents a promising avenue for mitigating these challenges in a cost-effective manner. However, existing methods often suffer from unsatisfactory generalization and unintended effects on non-edited samples. To overcome these limitations, we introduce a novel approach: Scalable Model Editing via Customized Expert Networks (SCEN), which is a two-stage continuous training paradigm. Specifically, in the first stage, we train lightweight expert networks individually for each piece of knowledge that needs to be updated. Subsequently, we train a corresponding indexing neuron for each expert to control the activation state of that expert. We conducted a series of experiments on the ZsRE and Hallucination benchmarks by tuning the advanced open-source LLM, Llama2, achieving state-of-the-art results compared to current mainstream methods. Our code is available at https://github.com/TAL-auroraX/SCEN.
Large Language Models (LLMs) have shown proficiency in question-answering tasks but often struggle to integrate real-time knowledge, leading to potentially outdated or inaccurate responses. This problem becomes even more challenging when dealing with multi-hop questions, since they require LLMs to update and integrate multiple knowledge pieces relevant to the questions. To tackle the problem, we propose the Retrieval-Augmented model Editing (RAE) framework for multi-hop question answering. RAE first retrieves edited facts and then refines the language model through in-context learning. Specifically, our retrieval approach, based on mutual information maximization, leverages the reasoning abilities of LLMs to identify chain facts that traditional similarity-based searches might miss. In addition, our framework includes a pruning strategy to eliminate redundant information from the retrieved facts, which enhances the editing accuracy and mitigates the hallucination problem. Our framework is supported by theoretical justification for its fact retrieval efficacy. Finally, comprehensive evaluation across various LLMs validates RAE's ability in providing accurate answers with updated knowledge. Our code is available at: https://github.com/sycny/RAE.
Large language models (LLMs) can make predictions using parametric knowledge--knowledge encoded in the model weights--or contextual knowledge--knowledge presented in the context. In many scenarios, a desirable behavior is that LLMs give precedence to contextual knowledge when it conflicts with the parametric knowledge, and fall back to using their parametric knowledge when the context is irrelevant. This enables updating and correcting the model's knowledge by in-context editing instead of retraining. Previous works have shown that LLMs are inclined to ignore contextual knowledge and fail to reliably fall back to parametric knowledge when presented with irrelevant context. In this work, we discover that, with proper prompting methods, instruction-finetuned LLMs can be highly controllable by contextual knowledge and robust to irrelevant context. Utilizing this feature, we propose EREN (Edit models by REading Notes) to improve the scalability and robustness of LLM editing. To better evaluate the robustness of model editors, we collect a new dataset, that contains irrelevant questions that are more challenging than the ones in existing datasets. Empirical results show that our method outperforms current state-of-the-art methods by a large margin. Unlike existing techniques, it can integrate knowledge from multiple edits, and correctly respond to syntactically similar but semantically unrelated inputs (and vice versa). The source code can be found at https://github.com/thunlp/EREN.
ROME and MEMIT are largely believed to be two different model editing algorithms, with the major difference between them being the ability to perform batched edits. In this paper, we unify these two algorithms under a single conceptual umbrella, optimizing for the same goal, which we call the preservation-memorization objective. ROME uses an equality constraint to optimize this objective to perform one edit at a time, whereas MEMIT employs a more flexible least-square constraint that allows for batched edits. We generalize ROME and enable batched editing with equality constraint in the form of EMMET - an Equality-constrained Mass Model Editing algorithm for Transformers, a new batched memory-editing algorithm. EMMET can perform batched-edits up to a batch-size of 10,000, with very similar performance to MEMIT across multiple dimensions. With the introduction of EMMET, we truly unify ROME and MEMIT and show that both algorithms are equivalent in terms of their optimization objective, their abilities (singular and batched editing), their model editing performance and their limitations.
Mainstream backdoor attack methods typically demand substantial tuning data for poisoning, limiting their practicality and potentially degrading the overall performance when applied to Large Language Models (LLMs). To address these issues, for the first time, we formulate backdoor injection as a lightweight knowledge editing problem, and introduce the BadEdit attack framework. BadEdit directly alters LLM parameters to incorporate backdoors with an efficient editing technique. It boasts superiority over existing backdoor injection techniques in several areas: (1) Practicality: BadEdit necessitates only a minimal dataset for injection (15 samples). (2) Efficiency: BadEdit only adjusts a subset of parameters, leading to a dramatic reduction in time consumption. (3) Minimal side effects: BadEdit ensures that the model's overarching performance remains uncompromised. (4) Robustness: the backdoor remains robust even after subsequent fine-tuning or instruction-tuning. Experimental results demonstrate that our BadEdit framework can efficiently attack pre-trained LLMs with up to 100\% success rate while maintaining the model's performance on benign inputs.
Text-to-image diffusion models suffer from the risk of generating outdated, copyrighted, incorrect, and biased content. While previous methods have mitigated the issues on a small scale, it is essential to handle them simultaneously in larger-scale real-world scenarios. We propose a two-stage method, Editing Massive Concepts In Diffusion Models (EMCID). The first stage performs memory optimization for each individual concept with dual self-distillation from text alignment loss and diffusion noise prediction loss. The second stage conducts massive concept editing with multi-layer, closed form model editing. We further propose a comprehensive benchmark, named ImageNet Concept Editing Benchmark (ICEB), for evaluating massive concept editing for T2I models with two subtasks, free-form prompts, massive concept categories, and extensive evaluation metrics. Extensive experiments conducted on our proposed benchmark and previous benchmarks demonstrate the superior scalability of EMCID for editing up to 1,000 concepts, providing a practical approach for fast adjustment and re-deployment of T2I diffusion models in real-world applications.
Large Language Models have revolutionized numerous tasks with their remarkable efficacy. However, editing these models, crucial for rectifying outdated or erroneous information, often leads to a complex issue known as the ripple effect in the hidden space. While difficult to detect, this effect can significantly impede the efficacy of model editing tasks and deteriorate model performance. This paper addresses this scientific challenge by proposing a novel evaluation methodology, Graphical Impact Evaluation(GIE), which quantitatively evaluates the adaptations of the model and the subsequent impact of editing. Furthermore, we introduce the Selective Impact Revision(SIR), a model editing method designed to mitigate this ripple effect. Our comprehensive evaluations reveal that the ripple effect in the hidden space is a significant issue in all current model editing methods. However, our proposed methods, GIE and SIR, effectively identify and alleviate this issue, contributing to the advancement of LLM editing techniques.
Interventions on model-internal states are fundamental operations in many areas of AI, including model editing, steering, robustness, and interpretability. To facilitate such research, we introduce $\textbf{pyvene}$, an open-source Python library that supports customizable interventions on a range of different PyTorch modules. $\textbf{pyvene}$ supports complex intervention schemes with an intuitive configuration format, and its interventions can be static or include trainable parameters. We show how $\textbf{pyvene}$ provides a unified and extensible framework for performing interventions on neural models and sharing the intervened upon models with others. We illustrate the power of the library via interpretability analyses using causal abstraction and knowledge localization. We publish our library through Python Package Index (PyPI) and provide code, documentation, and tutorials at https://github.com/stanfordnlp/pyvene.
Recent work using Rank-One Model Editing (ROME), a popular model editing method, has shown that there are certain facts that the algorithm is unable to edit without breaking the model. Such edits have previously been called disabling edits. These disabling edits cause immediate model collapse and limits the use of ROME for sequential editing. In this paper, we show that disabling edits are an artifact of irregularities in the implementation of ROME. With this paper, we provide a more stable implementation ROME, which we call r-ROME and show that model collapse is no longer observed when making large scale sequential edits with r-ROME, while further improving generalization and locality of model editing compared to the original implementation of ROME. We also provide a detailed mathematical explanation of the reason behind disabling edits.
As the typical retraining paradigm is unacceptably time- and resource-consuming, researchers are turning to model editing to find an effective way that supports both consecutive and batch scenarios to edit the model behavior directly. Despite all these practical expectations, existing model editing methods fail to realize all of them. Furthermore, the memory demands for such sequential model editing approaches tend to be prohibitive, frequently necessitating an external memory that grows incrementally over time. To cope with these challenges, we propose CoachHooK, a model editing method that simultaneously supports sequential and batch editing. CoachHooK is memory-friendly as it only needs a small amount of it to store several hook layers whose size remains unchanged over time. Experimental results demonstrate the superiority of our method over other batch-supportive model editing methods under both single-round and consecutive batch editing scenarios. Extensive analyses of CoachHooK have been conducted to verify the stability of our method over a number of consecutive steps.
Model editing has emerged as a cost-effective strategy to update knowledge stored in language models. However, model editing can have unintended consequences after edits are applied: information unrelated to the edits can also be changed, and other general behaviors of the model can be wrongly altered. In this work, we investigate how model editing methods unexpectedly amplify model biases post-edit. We introduce a novel benchmark dataset, Seesaw-CF, for measuring bias-related harms of model editing and conduct the first in-depth investigation of how different weight-editing methods impact model bias. Specifically, we focus on biases with respect to demographic attributes such as race, geographic origin, and gender, as well as qualitative flaws in long-form texts generated by edited language models. We find that edited models exhibit, to various degrees, more biased behavior as they become less confident in attributes for Asian, African, and South American subjects. Furthermore, edited models amplify sexism and xenophobia in text generations while remaining seemingly coherent and logical. Finally, editing facts about place of birth, country of citizenship, or gender have particularly negative effects on the model's knowledge about unrelated features like field of work.
Model editing aims to precisely alter the behaviors of large language models (LLMs) in relation to specific knowledge, while leaving unrelated knowledge intact. This approach has proven effective in addressing issues of hallucination and outdated information in LLMs. However, the potential of using model editing to modify knowledge in the medical field remains largely unexplored, even though resolving hallucination is a pressing need in this area. Our observations indicate that current methods face significant challenges in dealing with specialized and complex knowledge in medical domain. Therefore, we propose MedLaSA, a novel Layer-wise Scalable Adapter strategy for medical model editing. MedLaSA harnesses the strengths of both adding extra parameters and locate-then-edit methods for medical model editing. We utilize causal tracing to identify the association of knowledge in neurons across different layers, and generate a corresponding scale set from the association value for each piece of knowledge. Subsequently, we incorporate scalable adapters into the dense layers of LLMs. These adapters are assigned scaling values based on the corresponding specific knowledge, which allows for the adjustment of the adapter's weight and rank. The more similar the content, the more consistent the scale between them. This ensures precise editing of semantically identical knowledge while avoiding impact on unrelated knowledge. To evaluate the editing impact on the behaviours of LLMs, we propose two model editing studies for medical domain: (1) editing factual knowledge for medical specialization and (2) editing the explanatory ability for complex knowledge. We build two novel medical benchmarking datasets and introduce a series of challenging and comprehensive metrics. Extensive experiments on medical LLMs demonstrate the editing efficiency of MedLaSA, without affecting unrelated knowledge.
In this study, we tackle a growing concern around the safety and ethical use of large language models (LLMs). Despite their potential, these models can be tricked into producing harmful or unethical content through various sophisticated methods, including 'jailbreaking' techniques and targeted manipulation. Our work zeroes in on a specific issue: to what extent LLMs can be led astray by asking them to generate responses that are instruction-centric such as a pseudocode, a program or a software snippet as opposed to vanilla text. To investigate this question, we introduce TechHazardQA, a dataset containing complex queries which should be answered in both text and instruction-centric formats (e.g., pseudocodes), aimed at identifying triggers for unethical responses. We query a series of LLMs -- Llama-2-13b, Llama-2-7b, Mistral-V2 and Mistral 8X7B -- and ask them to generate both text and instruction-centric responses. For evaluation we report the harmfulness score metric as well as judgements from GPT-4 and humans. Overall, we observe that asking LLMs to produce instruction-centric responses enhances the unethical response generation by ~2-38% across the models. As an additional objective, we investigate the impact of model editing using the ROME technique, which further increases the propensity for generating undesirable content. In particular, asking edited LLMs to generate instruction-centric responses further increases the unethical response generation by ~3-16% across the different models.
Task embedding, a meta-learning technique that captures task-specific information, has gained popularity, especially in areas such as multi-task learning, model editing, and interpretability. However, it faces challenges with the emergence of prompt-guided Large Language Models (LLMs) operating in a gradient-free manner. Existing task embedding methods rely on fine-tuned, task-specific language models, which hinders the adaptability of task embeddings across diverse models, especially prompt-based LLMs. To hardness the potential of task embeddings in the era of LLMs, we propose a framework for unified task embeddings (FUTE), harmonizing task embeddings from various models, including smaller language models and LLMs with varied prompts, within a single vector space. Such uniformity enables comparison and analysis of similarities amongst different models, broadening the scope and utility of existing task embedding methods in multi-model scenarios, while maintaining their performance comparable to architecture-specific methods.